Rethinking Retinal Image Quality: Treating Quality Threshold as a Tunable Hyperparameter

Abstract

Assuming the robustness of a deep learning model to suboptimal images is a key consideration, we asked if there was any value in including training images of poor quality. In particular, should we treat the (quality) threshold at which a training image is either included or excluded as a tunable hyperparameter? To that end, we systematically examined the effect of including training images of varying quality on the test performance of a DL model in classifying the severity of diabetic retinopathy. We found that there was a unique combination of (categorical) quality labels or a Goldilocks (continuous) quality score that gave rise to optimal test performance on either high-quality or suboptimal images. The model trained exclusively on high-quality images yielded worse performance in all test scenarios than that trained on the optimally tuned training set which included images with some level of degradation.